SURF Final Report: Deformable Models for Segmentation

Methods for generating a closed geometric model from a point-sampled volume data set can greatly enhance visualization of noisy volume data collected from nondestructive sensing instruments such as MRI. Traditional methods attempt to perform segmentation automatically, by identifying boundary layers using a parametric snake. Essentially, an initial model, called a ‘seed’ is placed inside the item of interest in the volume data. Then, the model is deformed by a relaxation process based on forces applied to each vertex of the model. Eventually, the forces will balance out, and the model will converge to a certain shape. Unfortunately, due to high levels of noise and low resolution between slices, that is typical in most MRI scans, it is often extremely difficult to obtain good results. We attempt to cope with this problem by developing a hybrid system, whereby the user places down guidance points along the boundary to help guide the identification of the boundary layers. Furthermore, we introduce recent techniques in mesh operations and smoothing to maintain good geometric properties throughout the expansion phase.